A Study Of Adaptive Scale Tracking Based On Structured SVM

Visual object tracking is a hot research topic in the field of computer vision,which has been widely used in intelligent visual surveillance systems and human-computer interaction,and has important research significance.Although visual object tracking technology has achieved rapid development,it will be greatly impacted by some factors in actual application scenes,such as deformation,occlusion,illumination changes and complex background.So how to meet the real-time,accuracy,and robustness requirement of tracking algorithms is a problem to be solved.Struck is an excellent structured output tracking algorithm,which is more accurate than traditional algorithms.Besides,it can be adapted to the deformation of objects as well as robust when occlusion occurs.However,the real-time of the Struck algorithm needs to be improved and its tracking accuracy will decrease significantly when target scale changes.Therefore,combining with the Struck algorithm,this paper proposes a new adaptive scale algorithm based on structured SVM.The main work is as follows:(1)Aiming at the shortage of real-time performance of the Struck algorithm,this paper introduces the process of object location prediction.Then,the improved algorithm can estimate positions by discriminator from the sparse samples,which can decrease the searching scope and calculation cost.As a result,this search strategy improves the real-time performance greatly.(2)Aiming at the decrease of tracking accuracy in the Struck algorithm when object scale changes,this paper proposes a new method of adaptive scale tracking.Firstly,scale variable is added to structured SVM model.Then,the samples of different sizes can be obtained in the process of confirming object location.Finally,the scale change is calculated by discriminator.As a result,the tracking accuracy is improved.(3)In order to solve the problems of drift and degradation in long time tracking,this paper uses the method of on-line incremental learning and limits the number of support vectors.When the number exceeds the threshold,the support vectors which impact discriminator least will be removed,so that the real-time and accuracy of tracking algorithm will be enhanced further.Finally,the performance of Struck algorithm and the improved algorithm are compared by experiments.Result shows that the improved algorithm has a better performance while tracking different objects,and even better accuracy when the object scale changes.